The purpose of space surveillance is to classify and, if possible, assess the mission and performance capabilities of space objects. Historically, imaging techniques have obtained useful results. However, with the advances achieved in microtechnologies, small but highly functional satellites (largest dimension <1 m) are emerging that are hard to identify by imaging with large ground-based telescopes. The concept of using nonimaging measurements to obtain information is relatively new. In this paper, we present and discuss the performance of two techniques for classifying satellites based on spectral measurements. A distance-based classifier and a neural-net-based classifier are used to process both calibrated spectral data and features computed using these data. Neural networks are found to give better recognition results than the distance-based classifier, and once trained, this method is also faster. The average error rates for the distance-based method are greater than 30% when the inputs are the calibrated spectra, and 70% when using the central moments and the K-nearest-neighbors method. The best results are obtained for the neural network design, with the lowest class error rate at 0% for some satellites, the highest error rate at 30%, and an average error rate at 16%.